Issues

• cosmos3: Cosmos3-Edge reasoner (image + text) as a native family on Metal and CUDA

  ashhart/TensorFold#1269 · Filter by author drowzeys opened on Oct 25, 2026
  1

• CUDA kernel contract: the Qwen 3.5 decode attention cases launch tf_q35_attn_partial and tf_q35_attn_partial8 at 192 threads, past their __launch_bounds__(128), so every gate fails with CUDA_ERROR_INVALID_VALUE

  ashhart/TensorFold#1267 · Filter by author Bizuayeu opened on Oct 25, 2026
  1

• Drama 3: native Metal support for the pinned NVFP4 pruned checkpoint

  ashhart/TensorFold#1266 · Filter by author jerryfane opened on Oct 25, 2026
  4

• CUDA kernel contract on GB10: tf_w4a8_prefill hits a misaligned address, so kernel-contract loses its context and ends without a SUMMARY since 1.1.4

  ashhart/TensorFold#1264 · Filter by author Bizuayeu opened on Oct 24, 2026
  1

• Flash Next: a document re-sent inside a user message re-prefills all 128k tokens (a prefix-cache limitation)

  ashhart/TensorFold#1262 · Filter by author devopsfortunato opened on Oct 24, 2026

• App: release beta.10 with TensorFold 1.1.4

  ashhart/TensorFold#1253 · Filter by author ashhart opened on Oct 24, 2026
  3

• Ling-3.0-tiny on a Qualcomm board: a Ling family on Metal first, then an ARM CPU backend

  ashhart/TensorFold#1247 · Filter by author ThinkOffApp opened on Oct 24, 2026

• Qwen4-27B (dense) on native CUDA, GB10: shape parameterization ahead of the checkpoint (RFC)

  ashhart/TensorFold#1244 · Filter by author jschmied opened on Oct 24, 2026
  3

• Non-streaming reply puts <tool_call> markup in content when max_tokens cuts a call (streaming returns the call)

  ashhart/TensorFold#1238 · Filter by author laidick opened on Oct 24, 2026
  3

• Tensor parallel across Metal and CUDA ranks (Mac + NVIDIA): tracker

  ashhart/TensorFold#1236 · Filter by author ThinkOffApp opened on Oct 24, 2026
  2

• Gemma 4 E4B (NVIDIA Nemotron 3.5 Bolt E4B): per-layer inputs and shared KV, Intel Arc first

  ashhart/TensorFold#1222 · Filter by author Mirrdhyn opened on Oct 23, 2026
  4

• RoCE transport above two ranks: roce.zig:148 refuses every world but 2, so four GB10s fall back to NCCL (rebased 2..16-rank branch, gate question)

  ashhart/TensorFold#1209 · Filter by author flashosophy opened on Oct 23, 2026

• Studio: carry #941 to #944 onto main as reviewed increments

  ashhart/TensorFold#1204 · Filter by author ashhart opened on Oct 22, 2026
  1

• Proposal: Qwen3.8-27B follow-up turns resume from slot 0 (Metal, small-memory Macs)

  ashhart/TensorFold#1197 · Filter by author RS91Dev opened on Oct 22, 2026
  1

• Decisions serving for qwen27 (Ornith-1.5-122B): score/decide passes, JointHead scoring, a free decision slot, GPU class profiling

  ashhart/TensorFold#1190 · Filter by author grantoverton opened on Oct 22, 2026

• New family: Recursive Flint (40B-A2.5B MoE) on Metal

  ashhart/TensorFold#1175 · Filter by author BobClawblaw opened on Oct 21, 2026

• CUDA: Qwen3.5/3.6 attention caches on the core KV pool (#781 follow-up)

  ashhart/TensorFold#1156 · Filter by author jschmied opened on Oct 21, 2026
  5

• Proposal: dense BF16 Qwen3.5 CUDA scoring

  ashhart/TensorFold#1136 · Filter by author akol1 opened on Oct 21, 2026

• Flash Next 4-bit doesn't fit natively on a 128 GB Mac at default limits. Any plans for --ple-on-ssd on Metal?

  ashhart/TensorFold#1039 · Filter by author hudsonwa opened on Oct 18, 2026
  3

• Proposal: several models on one Mac under one memory budget

  ashhart/TensorFold#1027 · Filter by author cwschroeder opened on Oct 18, 2026
  1

• Proposal: TensorFold Notes, memory inside the server (for 1.1.0)

  ashhart/TensorFold#1012 · Filter by author ThinkOffApp opened on Oct 17, 2026
  6

• Benchmarks: accept ROCm receipts on the board (Strix Halo)

  ashhart/TensorFold#1001 · Filter by author ThinkOffApp opened on Oct 17, 2026
  6

• GLM-5.3-Flash TP=2 (0.6.0 lineage): with uptime a 13-token prefill reaches 18.8 s and decode falls 40 to 14 tok/s, an idle rank 1 reads 96% GPU, and a restart clears both

  ashhart/TensorFold#993 · Filter by author ss-cong opened on Oct 17, 2026
  18

• Nemotron 3.5 Super VL on two GB10s: TP=2 serial decode is no faster than TP=1 here, and what TP=3/4 would need

  ashhart/TensorFold#981 · Filter by author easthambuilds opened on Oct 17, 2026
  5

• Ling-3.1-flash on CUDA: a shared KDA kernel first, then the family at TP=1 and TP=2 (EXL3)

  ashhart/TensorFold#850 · Filter by author jschmied opened on Oct 14, 2026
  6